Resource searching method and device, equipment, storage medium and program product

Abstract

The invention provides a resource search method and device, equipment, a storage medium and a program product, and the method comprises the steps: carrying out the intention recognition of a resource search request inputted by a user based on a pre-trained large language model, and obtaining a request resource entity and a request resource type; determining a plurality of target search engines from a plurality of search engines based on the request resource type, and expanding search terms based on the request resource type; calling each target search engine for searching based on the request resource entity and the expanded search term; and verifying a search result returned by the target search engine, and generating a response of the resource search request based on the verified search result. Multiple search engines are used for parallel retrieval, so that the coverage degree and the search success rate are improved; by automatically verifying the search result, the search result filtering is realized, the user operation is simplified, and the convenience of resource downloading is improved.

Description

Technical Field

[0001] This application relates to the field of large language model technology, and in particular to a resource search method, apparatus, device, storage medium and program product. Background Technology

[0002] With the deep integration of the digital economy and internet technology, users' demand for downloading digital resources in learning, work, and production scenarios continues to grow.

[0003] Traditional digital resource search methods rely on users inputting precise keywords, placing high demands on user skills. Against this backdrop, Large Language Models (LLM) have been applied to resource search to leverage their powerful natural language understanding capabilities and lower the input barrier for users. However, existing LLM-based resource searches often use a single search engine, resulting in limited resource coverage and search efficiency. Furthermore, LLMs typically return multiple search results, such as multiple web pages, requiring users to sift through these results to find downloadable links—a time-consuming and laborious process.

[0004] Therefore, there is an urgent need to provide a more convenient resource search solution based on a large language model that supports multi-search engine collaboration in order to improve the success rate and efficiency of resource search. Summary of the Invention

[0005] This application provides a resource search method, apparatus, device, storage medium, and program product. It utilizes a large model to achieve accurate intent recognition, significantly reducing the search threshold; it matches the identified resource type with multiple search engines to improve search efficiency and success rate; and it automatically verifies search results, eliminating the need for users to verify search results and generating request responses using verified search results, thereby improving the convenience of resource search and download.

[0006] In a first aspect, embodiments of this application provide a resource search method, comprising: performing intent recognition on a user-inputted resource search request based on a pre-trained large language model to obtain a requested resource entity and a requested resource type; determining multiple target search engines from multiple search engines based on the requested resource type, and expanding search terms based on the requested resource type; invoking each target search engine to perform a search based on the requested resource entity and the expanded search terms; verifying the search results returned by the target search engines, and generating a response to the resource search request based on the verified search results.

[0007] Secondly, embodiments of this application provide a resource search apparatus comprising: an intent recognition module, used to recognize the intent of a user-inputted resource search request based on a pre-trained large language model, and obtain the requested resource entity and the requested resource type; a search engine determination module, used to determine multiple target search engines from multiple search engines based on the requested resource type, and to expand search terms based on the requested resource type; a search module, used to call each target search engine to perform a search based on the requested resource entity and the expanded search terms; and a verification and response generation module, used to verify the search results returned by the target search engines, and to generate a response to the resource search request based on the verified search results.

[0008] Thirdly, embodiments of this application provide an electronic device, including a memory and a processor; the memory stores computer-executable instructions; the processor executes the computer-executable instructions stored in the memory, causing the processor to perform the method and / or various implementations provided in the first aspect of this application.

[0009] Fourthly, embodiments of this application provide a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement the method and / or various implementation methods provided in the first aspect of this application.

[0010] Fifthly, embodiments of this application provide a computer program product, including a computer program that, when executed by a processor, implements the method and / or various implementation methods provided in the first aspect of this application.

[0011] The resource search method, apparatus, device, storage medium, and program products provided in this application utilize a large language model to identify the intent of resource search requests, effectively lowering the search threshold. Users are not required to provide precise search keywords, and the system supports colloquial and ambiguous expressions of needs, improving the convenience of search operations. Simultaneously, the large model can accurately uncover users' potential needs, improving the accuracy of intent recognition. Based on the resource type obtained through intent recognition, multiple search engines are automatically matched for retrieval, increasing coverage compared to single-engine searches. Furthermore, matching search engines by resource type reduces invalid search engine calls, improving search efficiency. By automatically verifying search results and eliminating invalid results, a request response is generated only based on verified and highly reliable search results, saving users the cost of manual verification and further improving the convenience of search operations. Attached Figure Description

[0012] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0013] Figure 1 An application scenario diagram provided for an embodiment of this application;

[0014] Figure 2 A flowchart illustrating the resource search method provided in this application embodiment;

[0015] Figure 3 A flowchart illustrating a resource search method provided in another embodiment of this application;

[0016] Figure 4 A flowchart illustrating a resource search method provided in another embodiment of this application;

[0017] Figure 5 A schematic diagram of the structure of the resource search device provided in the embodiments of this application;

[0018] Figure 6 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application.

[0019] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to particular embodiments. Detailed Implementation

[0020] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0021] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, use and processing of the relevant data must comply with relevant laws, regulations and standards, and corresponding operation entry points are provided for users to choose to authorize or refuse.

[0022] The technical solution of this application and how the technical solution of this application solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will be described below with reference to the accompanying drawings.

[0023] The method provided in this application can be widely applied to various resource download scenarios, covering a variety of resources such as documents, images, videos, code, audio, and compressed files.

[0024] For example, a user can initiate a resource search request through the browser's input box, and the browser can call this method to process the input resource search request and return the search results.

[0025] In addition to browsers, users can also initiate resource search requests through the interfaces of applications (APPs) installed on their devices, as well as mini-programs within those applications, to perform resource searches.

[0026] For example, Figure 1 An application scenario diagram provided for an embodiment of this application, such as... Figure 1 As shown, the user terminal has a resource search assistant installed. This assistant can exist as a standalone app, plugin, or embedded function in third-party software. Users input natural language search queries, such as "download Python's TensorFlow documentation," into the resource search assistant's dialog box based on their resource needs. The resource search assistant then performs intent recognition on the search query, identifying "download TensorFlow documentation," and invokes a search engine to retrieve the search results. After obtaining the search results, the resource search assistant generates a response based on those results.

[0027] The response typically includes a script, such as "TensorFlow documentation has been found for you. Three results have been filtered out. You can view the details and download it." It may also include multiple information cards, each clearly displaying the core information of a single search result, including: title (e.g., title 1, title 2, and title 3), cover image (e.g., img1, img2, and img3), and summary information (…). Figure 1 (Using ellipses to indicate the location). Each card can also display attribute information at the bottom, such as document size, resource source, download conditions, and buttons like "Download" and "View," to facilitate user operation. Users can click the "View" button or the displayed information card to view the details page of the corresponding search result.

[0028] The resource search assistant uses a large language model to identify the intent of search queries and generate search result summaries.

[0029] Figure 1 The resource search scenarios and result display methods shown are merely examples and are not intended to limit this application. In some embodiments, the resource search assistant may display search results in other ways or initiate resource search requests through other carriers, such as the aforementioned browser.

[0030] Both browsers and resource search assistants have the following problems when using large language models for resource searching: they usually call a single search engine, resulting in incomplete resource coverage and low retrieval efficiency; they directly return multiple top-ranked search results based on the search engine's ranking of search results, lacking a verification mechanism for search results, which requires users to manually verify whether the links are valid, whether payment is required, and whether the information is accurate, increasing the user's operational costs.

[0031] To address the aforementioned issues, this application provides a resource search method that identifies search intent through a large language model to obtain the resources the user wishes to download and their types. It then performs parallel searches using multiple search engines adapted to different resource types, improving resource coverage and search success rate. Simultaneously, it introduces a search result verification mechanism to verify the accuracy, validity, reliability, and accessibility of search results, automatically eliminating erroneous, invalid, or high-barrier-to-access search results. This eliminates the need for user verification, improving convenience and reliability.

[0032] Figure 2 This is a flowchart illustrating the resource search method provided in an embodiment of this application. This resource search method can be executed by any device or module with corresponding data processing capabilities, such as a large language model, a server-side application for a large language model, or a terminal integrating a large language model. Figure 2 As shown, the resource search method includes:

[0033] Step S201: Based on the pre-trained large language model, perform intent recognition on the user's input resource search request to obtain the requested resource entity and the requested resource type.

[0034] Here, the requested resource entity is the entity corresponding to the resource requested by the user, which can specifically be an entity extracted from the resource search request that represents the name of the resource the user wants to download. The requested resource type is the type of resource requested by the user.

[0035] Intent recognition specifically refers to identifying the resources that a user wants to search for or download and their types, thereby obtaining the requested resource entity and the requested resource type.

[0036] The requested resource entity can be an entity corresponding to a resource name such as software name, paper title, or document name.

[0037] A pre-trained large language model can be further fine-tuned based on this pre-trained model. Pre-training specifically refers to training the large language model on large-scale text data beforehand, enabling it to possess natural language understanding and generation capabilities. Then, based on resource search needs, specialized training data for resource search is used to fine-tune the pre-trained large language model, giving it intent recognition capabilities. If subsequent steps also rely on the large language model, training data can be used to equip it with the corresponding capabilities for those steps, such as expanding search terms, invoking search engines, and verifying search results.

[0038] Resource search requests can be search statements described in natural language by the user, or they can include multiple keywords, or they can be structured data.

[0039] For example, users can input their download requests using modalities such as voice or images. The large language model obtains the text modal input information, i.e., the resource search request, through methods such as speech-to-text conversion and text recognition in images.

[0040] Intent-based prompts can be pre-input into a large language model, enabling the model to identify the intent of subsequent resource search requests under the guidance of these prompts, thus obtaining the requested resource entity and the requested resource type. The intent-based prompts guide the large language model to extract the entity representing the resource name from the input resource search request, thereby obtaining the resource entity the user is requesting to download and determining the type of resource requested.

[0041] The intent recognition prompts were designed by relevant personnel based on the core needs of resource download scenarios and generated after being corrected using labeled samples corresponding to resource download scenarios.

[0042] Intent recognition prompts can also include information about the output format of the intent recognition result (including the requested resource entity and the requested resource type), such as JSON (JavaScript Object Notation) format, XML (Extensible Markup Language) format, etc.

[0043] For example, the intent recognition result can be in JSON format. Taking a resource search request as "download the TensorFlow documentation for Python" as an example, the intent recognition result obtained by the large language model after performing intent recognition on the user's resource search request can be:

[0044] {

[0045] "keyword_title": "TensorFlow documentation",

[0046] "type": "Python documentation"

[0047] }

[0048] Here, keyword_title represents the requested resource entity, and type represents the requested resource type.

[0049] The intent recognition prompt may also include one or more labeled examples, each of which includes a corresponding resource search request and an output example.

[0050] In intent recognition, the large language model can also be based on the stored temporal sequence of the user's search intent. Specifically, the temporal sequence of the user's search intent and the current input resource search request are input into the large language model, enabling the model to more accurately identify the search intent of the input resource search request based on the user's search intent temporal sequence, and obtain the requested resource entity and requested resource type.

[0051] Search intent time series is a record of the resources and their types that a user requested to access or download each time they searched for resources in a historical period. Unlike typical user search history logs, search intent time series records structured data with timestamps and intent identification related to user search behavior. It reflects the changing characteristics of structured data consisting of the user's search intent, i.e., the requested resource entity and the requested resource type, over time.

[0052] Search intent time series includes intent recognition results corresponding to historical search behaviors. For each user, the large language model performs intent recognition on the user's input resource search request, and generates a response to the resource search request through subsequent steps based on the intent recognition result. If the user downloads the resource through the response, the intent recognition result is stored at the end of the search intent time series, thus continuously updating the user's search intent time series. It is also possible to limit the upper limit of intent recognition results stored in each user's search intent time series; when this limit is reached, new data overwrites old data, ensuring that only the user's most recent intent recognition results are retained in the time series.

[0053] Step S202: Based on the requested resource type, determine multiple target search engines from multiple search engines, and expand the search terms based on the requested resource type.

[0054] After determining the requested resource type through intent recognition, multiple search engines corresponding to the requested resource type are identified from the pre-stored correspondence between the resource type and the search engine. Each search engine corresponding to the requested resource type can be directly identified as the target search engine.

[0055] A large number of search tests of various types of resources can be conducted in advance through various search engines. Based on the quantitative evaluation indicators of the search results, the adaptability of each search engine to each type of resource can be determined, and the aforementioned correspondence can be obtained.

[0056] In this correspondence, each resource type corresponds to at least two search engines.

[0057] Optionally, based on the requested resource type, multiple target search engines are determined from multiple search engines, including: obtaining multiple candidate search engines corresponding to the requested resource type; obtaining the frequency of users downloading resources of the requested resource type through the search results returned by each candidate search engine; and determining multiple target search engines based on the frequency.

[0058] The multiple candidate search engines corresponding to the requested resource type are the search engines corresponding to the requested resource type recorded in the aforementioned correspondence.

[0059] You can select multiple candidate search engines that frequently download the requested resource type based on the returned search results as multiple target search engines.

[0060] A table can be used to record the frequency of users downloading resources of various types through different search engines, and this table can be dynamically updated based on user download activity. For example, after generating a response to a resource search request, such as a download link, in subsequent steps, if a user downloads the corresponding resource information through that download link, the download count for that resource type by the search engine corresponding to that download link is incremented by 1. Based on the statistical count of download counts for each resource type by search engines, the frequency of each search engine for each resource type can be calculated.

[0061] By integrating resource types and users' historical download frequency, the search engine to be used for this resource search is determined. While ensuring that the search engine is adapted to the current resource download scenario, the search results are made more in line with the user's search habits and preferences, thereby improving the matching degree between search results and users.

[0062] The requested resource entity obtained through intent recognition can accurately reflect a user's download needs. However, in some scenarios, such as complex network resource environments, different types of resource names may be duplicated. Searching solely based on the requested resource entity may not allow the search engine to accurately determine the user's desired resource, resulting in inaccurate search results. To address this issue, this step implements search term expansion based on the requested resource type obtained through intent recognition.

[0063] When expanding search terms based on the requested resource type, semantic expansion can be performed on the requested resource type to obtain expanded search terms. Taking the resource search request as "download Python's TensorFlow documentation" as an example, the requested resource type is "Python documentation", and the expanded search terms can include "Python PDF documentation" and "Python technical documentation".

[0064] When expanding search terms, you can also consider using aliases, versions, etc. of the requested resource entity, that is, expanding the aliases, versions, etc. of the requested resource entity into search terms.

[0065] Search term expansion can be performed using a large language model, which involves semantically expanding the requested resource type to obtain expanded search terms. Alternatively, this step can be executed by the large language model to achieve matching with the target search engine and expand the search terms.

[0066] Step S203: Based on the requested resource entity and the expanded search terms, call each target search engine to perform a search.

[0067] For each target search engine, a search is initiated using the requested resource entity and expanded search terms as keywords to achieve parallel retrieval across multiple search engines and collect the search results returned by each target search engine.

[0068] Multiple target search engine APIs can be called in parallel. By setting a timeout mechanism, excessively long wait times for search results can be avoided. For example, the maximum wait time can be set to 3 seconds. If a target search engine does not return a search result within 3 seconds, the collection of search results from that target search engine will be actively terminated. If a search result is returned within 3 seconds, it will be received normally and subsequent verification will be performed.

[0069] In some embodiments, a search result is collected and then verified. If the first verified search result appears, the collection is terminated, i.e., no more search results are collected, in order to improve resource acquisition efficiency and reduce the cost of processing invalid data.

[0070] This step can be performed by a large language model, which can initiate API call requests to each target search engine based on the requested resource entity and expanded search terms, and collect the search results returned by each target search engine.

[0071] Step S204: Verify the search results returned by the target search engine, and generate a response to the resource search request based on the verified search results.

[0072] The response to a resource search request can be a structured result generated by sorting the verified search results according to preset sorting rules, such as sorting by the order of return time or by the publication time of the search results.

[0073] The response to a resource search request can be generated based on the first verified search result, such as the download link corresponding to the first verified search result.

[0074] For each search result returned by each target search engine, the search result is verified to check the accuracy and downloadability of the resources within the search results, as well as to identify any risks. If the results are inaccurate, undownloadable, or pose a risk, the verification fails, and subsequent collected search results are then verified.

[0075] It can extract attribute information from each search result, including one or more of the following: title, summary, download link, format, size, source, update time, and additional information. The additional information may include resource description, user rating, download volume, etc. The search result is then verified based on the extracted attributes.

[0076] Validation of search results can include intent validation, which verifies whether the resource corresponding to the search result matches the identified search intent. If the type of the resource in the search result does not match the type of the requested resource, the intent validation fails.

[0077] Optionally, the search results returned by the target search engine are validated, including: performing intent validation on the search results returned by the target search engine; intent validation is used to verify whether the search results match the type of the requested resource.

[0078] The type of resource corresponding to the search result can be determined by the title, summary, description, and resource format of the search result. If the type is consistent with the type of the requested resource, the intent verification passes; otherwise, the intent verification fails.

[0079] For example, if the resource search request is "open source code in XX paper", and the resource corresponding to the search result is "XX paper", which is of paper type, while the requested resource type is code, then the intent verification will fail.

[0080] Validation of search results can also include availability verification, which verifies whether the resource corresponding to the search result is available and downloadable. If the search result contains a download link for the resource, the validity of the link is verified. This can be done by simulating an access to the download link via an HTTP request to determine if there are any access errors, such as "404 Page Not Found," "503 Server Error," or expired link. If any such errors exist, the availability verification fails. Availability verification is also used to verify whether there are any barriers to downloading the resource, such as payment, membership, registration, or login. If so, the availability verification fails.

[0081] Based on the format, size, and update time of resources in the search results, it can be determined whether a resource is complete, and if not, it will fail the availability verification.

[0082] Verification of search results can also include security verification, used to verify whether the resources corresponding to the search results contain viruses, Trojans, etc.

[0083] The verification of search results may include one or more of the aforementioned verifications. If any one of the verifications fails, the verification of the search result fails.

[0084] Optionally, the search results returned by the target search engine may be validated, including: usability validation of the search results returned by the target search engine.

[0085] Optionally, the search results returned by the target search engine are validated, including: intent validation and availability validation of the search results returned by the target search engine; intent validation is used to verify whether the search results match the type of the requested resource.

[0086] Validating search results includes at least intent validation and usability validation. These validations can be performed in parallel or sequentially, for example, performing intent validation first and then usability validation, or vice versa.

[0087] By verifying intent and usability, the accuracy and usability of search results are ensured, significantly reducing the cost of ineffective user operations. This allows users to quickly obtain resources that meet their needs and can be used directly without additional screening, thus improving the user's resource search experience.

[0088] After obtaining the first verified search result, the system can continue to verify other search results, returning multiple responses for the user to choose from. Alternatively, it can stop verifying and collecting other search results, returning only the response for the first verified search result. The response to the search result is a structured download-oriented feedback for the user. This can be a download link that the user can directly act on based on the search result, or the download link along with some supplementary text (such as the title of the search result, operation prompts, etc.), or a structured page that displays relevant content from the search result using a preset style, such as title, source, number of downloads, rating, and file size.

[0089] When generating a response to a resource search request based on a verified search result, specifically, a resource download link can be generated and returned based on the first verified search result. For example, a response containing the resource download link can be output in a dialog box.

[0090] When generating a response to a resource search request based on verified search results, the response can specifically generate and return the resource download link corresponding to each verified search result.

[0091] In addition to the aforementioned resource download links, the response to a resource search request can also be a card, which displays some attribute information of the verified search results, such as title, summary, source, size, etc.

[0092] The resource search method provided in this embodiment utilizes a large language model to identify the intent of resource search requests, effectively lowering the search threshold. It eliminates the need for users to provide precise search keywords, supports colloquial and vague expressions of needs, and improves the convenience of the search operation. Simultaneously, the large model can accurately uncover users' potential needs, improving the accuracy of intent recognition. Based on the resource type obtained through intent recognition, multiple search engines are automatically matched for retrieval, increasing coverage compared to single-engine searches. Furthermore, matching search engines by resource type reduces invalid search engine calls, improving search efficiency. By automatically verifying search results and eliminating invalid results, a request response is generated only based on verified and highly reliable search results, saving users the cost of manual verification and further improving the convenience of the search operation.

[0093] Optionally, the intent recognition prompt includes multiple resource types to guide the large language model to extract the requested resource entity of the input resource search request, and to determine the requested resource type from multiple resource types; the method further includes: if there is no search result that passes intent verification, then replacing or deleting the requested resource type from the multiple resource types in the intent recognition prompt to obtain an updated intent recognition prompt, and returning to the step of performing intent recognition on the user's input resource search request, wherein the prompt used during intent recognition is the updated intent recognition prompt.

[0094] Optionally, based on a pre-trained large language model, intent recognition is performed on the user-inputted resource search request to obtain the requested resource entity and the requested resource type. This includes: extracting entities from the user-inputted resource search request based on the pre-trained large language model to obtain the requested resource entity and resource type entity; and, if the resource type entity includes polysemous words, obtaining the context of the resource search request; extracting scene features of the context; and disambiguating the resource type entity based on the scene features and a second knowledge graph to obtain the requested resource type. The second knowledge graph is used to represent the association between each resource type entity and various scene features.

[0095] Here, the resource type entity is the entity that represents the resource type. The context of the resource search request refers to the information provided to the user before and after entering the resource search request, such as the dialogue information before and after the resource search request is entered in a multi-turn dialogue scenario. The scenario features of the context are used to characterize the scenario of the resource search request.

[0096] Scene features can be extracted from the context by performing keyword extraction and topic analysis.

[0097] To improve the accuracy of resource type identification, large language models can also use a second knowledge graph to disambiguate ambiguous resource type entities.

[0098] Taking the resource type entity "Python" as an example, its meaning includes both language and software.

[0099] The second knowledge graph uses each resource type and scene feature as nodes. The edges connecting the nodes and their attributes are used to represent the matching degree of a corresponding set of resource types and scene features. Nodes corresponding to different resource types can also be connected by edges, and the attributes of these edges are used to represent the similarity between different resource types.

[0100] Scene features extracted from the context can be matched with scene features in the second knowledge graph. If a target scene feature that matches the scene feature extracted from the context exists, the resource type corresponding to the target scene feature is determined to be the requested resource type.

[0101] If multiple target scene features exist, the requested resource type is determined based on the weighted result of the resource types corresponding to the multiple target scene features. The weights used in the weighting are the attributes of the corresponding edges in the second knowledge graph.

[0102] By using context and knowledge graphs to disambiguate keywords in resource types, the accuracy of disambiguation is improved, which in turn improves the accuracy of intent recognition and reduces invalid searches caused by misjudgment of requested resource types.

[0103] Optionally, based on a pre-trained large language model, intent recognition is performed on the resource search request input by the user to obtain the requested resource entity and the requested resource type. This includes: obtaining the user's search intent time-series data; the search intent time-series data is used to record the requested resource type and its associated data corresponding to each search in the user's historical time; the search intent time-series data and the resource search request are input into the large language model, so that the large language model performs intent recognition on the resource search request based on the search intent time-series data to obtain the requested resource entity and the requested resource type.

[0104] Search intent time-series data is constructed for a single user and can be sorted in chronological order, providing structured data related to search intent.

[0105] The associated data for the requested resource type may include time, the input resource search request, the extracted requested resource entity, and may also include a summary of the downloaded resource.

[0106] When identifying the intent of a current resource search request, the large language model can identify the current search intent based on the user's historical intent recorded in the search intent time series data over a recent period.

[0107] When the user's current search request is too vague, combining the user's search intent time series data can help the large language model to infer the user's current search intent. By using history as a benchmark for inference, model illusions are reduced and the accuracy of resource search is improved.

[0108] After inputting the temporal data of search intent and the resource search request into the large language model, the large language model performs intent recognition on the resource search request under the guidance of graph recognition prompts. Specifically, under the guidance of graph recognition prompts, the following steps can be performed: extracting the resource search request entity input by the user to obtain the requested resource entity and the resource type entity; obtaining the context of the resource search request when the resource type entity includes polysemous words; extracting the scene features of the context; and disambiguating the resource type entity based on the scene features and the second knowledge graph to obtain the requested resource type. The second knowledge graph is used to represent the association between each resource type entity and various scene features.

[0109] Figure 3 This is a flowchart illustrating a resource search method provided in another embodiment of this application. Figure 2 Based on the embodiments, the resource search method is described in detail, such as... Figure 3 As shown, this resource search method may specifically include the following steps:

[0110] Step S301: Obtain the resource search request described in natural language by the user.

[0111] Step S302: Based on the pre-trained large language model and guided by intent recognition prompts, the intent of the resource search request is recognized to obtain the requested resource entity and the requested resource type.

[0112] Among them, intent recognition prompts are used to guide the large language model to extract the requested resource entity of the input resource search request, and to determine the requested resource type from multiple resource types.

[0113] Intent recognition prompts can include role definitions, task objectives, processing logic, and output format requirements.

[0114] Among them, the role definition is used to define the role that the large language model will play, such as "you are a resource download assistant" or "you are a resource download intent recognition assistant".

[0115] The task objective is used to inform the large language model of the specific task to be completed, including input and output. For example, "Based on the 'original keywords' provided by the user, analyze and extract the core intent of the resource demand, including the specific name and type of the resource." Here, the "original keywords" are the resource search request input by the user or the result of word segmentation processing of the request.

[0116] The processing logic details the rules and steps for analyzing user resource needs to make intent recognition more accurate. It may include classification rules and information extraction rules. Classification rules guide the large language model to determine whether the user's download needs are related to the preset resource types. If they are related, they are classified into the corresponding type tags. Information extraction rules are used to extract the core resource subject, i.e. the requested resource entity, from the original keywords.

[0117] For example, the processing logic in the intent recognition prompt can be: "Determine whether the user's intent is related to Python documentation, C++ documentation, JPG images, etc.: If related, return the relevant type. Core resource IP: Extract the core theme of the original keywords (such as document name, image name, learning material name)", where the core resource IP is the aforementioned requested resource entity.

[0118] The output format requirements are used to standardize the output format of the large language model. For example, it can be "Please strictly use this JSON format: {"keyword_title":"","type":""}".

[0119] For example, large language models can be Longformer (Long Range Attention Transformer), BERT (Bidirectional Encoder Representations from Transformers), GPT (Generative Pre-trained Transformer) series, etc.

[0120] Step S303: Based on the requested resource type, determine multiple target search engines from multiple search engines, and expand the search terms based on the requested resource type.

[0121] Step S304: Based on the requested resource entity and the expanded search terms, call each target search engine to perform a search, and collect the search results returned by the target search engines.

[0122] Step S305: Verify the usability of the search results returned by the target search engine.

[0123] Availability verification is used to verify the availability of resources in search results, including whether there are any barriers to obtaining them and whether the resources are valid.

[0124] Since the download methods for resources in search results are diverse, usability verification can be performed using an adaptation method based on the different download methods.

[0125] Some search results provide static download links, which can be directly verified to confirm their validity. Other search results provide resource access via dynamically loaded content, such as JavaScript. For these latter results, usability verification requires parsing the dynamically loaded content first to retrieve the hidden resource links.

[0126] Optionally, usability verification includes resource access threshold verification and validity verification. Usability verification is performed on the search results returned by the target search engine, including: verifying the resource access threshold of the search results; and verifying the validity of the links in the search results if the resource access entry in the search results is a static link.

[0127] Static links, in contrast to dynamically loaded content, refer to resource access addresses that are directly presented and can be accessed without additional loading.

[0128] The link can be accessed by simulating a request, and its validity can be verified by judging the link status based on the response. If the response is a normal status code and the metadata of the resource file can be obtained, the link is considered valid, i.e., it passes the validity verification. If the response is an abnormal status code, such as 404, 503, timeout, or redirection to an advertisement page, the link is considered invalid, i.e., it fails the validity verification.

[0129] In addition to validating the validity of the links, it is also necessary to verify the resource acquisition threshold of the verified search results to determine whether payment, registration, login, etc. are required to obtain the resource from the search results. If so, the search results are determined to have failed the resource acquisition threshold verification; otherwise, the resource acquisition threshold verification is passed.

[0130] In some embodiments, if login is required, user-authorized login information can be obtained and used to log in. If login is successful, the requirement to log in does not affect the resource access threshold verification result. Login information can be pre-stored or obtained through real-time interaction with the user. If login is not possible with payment, registration, or user-authorized login information, it is determined that the search results have failed the resource access threshold verification.

[0131] By directly verifying the validity of search results containing static links and performing threshold verification on search results, the verification efficiency is high. It effectively eliminates search results with invalid links or those requiring payment or registration, enabling the return of directly accessible resources and improving the convenience of resource downloads for users.

[0132] Optionally, the availability verification of the search results returned by the target search engine also includes: if the resource access entry in the search results is hidden in dynamically loaded content, then the dynamically loaded content is parsed to extract the hidden resource links contained in the dynamically loaded content; and the validity of the hidden resource links is verified.

[0133] Dynamically loaded content refers to search results, such as content on a webpage, that is not fully displayed upon initial loading but requires user interaction or automatic script execution before it can be loaded and displayed. This is often used to hide resource download links, such as links that only load when scrolling to the bottom of the page.

[0134] By examining the source code of search results, such as web pages, you can determine the type of search result. Specifically, you can determine whether the entry point for the requested resource in the search result is a static link or dynamically loaded content. If the source code contains a complete link to the requested resource, it's a static link; if no complete link is found, and only a placeholder exists, it's dynamically loaded content. You can also verify whether the search results download resources via dynamically loaded content by disabling JavaScript.

[0135] For dynamically loaded content, you can parse it, for example, by parsing the DOM (Document Object Model), and extract the hidden resource links of the dynamically loaded content.

[0136] Once the hidden resource link is extracted, it can be verified using the same methods as for verifying the validity of static links.

[0137] For search results containing dynamically loaded content, the parsing of dynamically loaded content breaks through the limitation of hidden resource entry points in dynamic content, improves the comprehensiveness of availability verification, avoids unusable resources from entering subsequent processes due to differences in resource entry point forms, and prevents valid resources from being lost because they cannot be verified.

[0138] Optionally, the search results are subjected to resource acquisition threshold verification, including: when the links in the search results are valid, simulating the resource acquisition process in the search results to detect whether there are payment, registration or login steps, and obtaining the verification result of the resource acquisition threshold verification.

[0139] This tool can leverage browser automation to reproduce the entire user experience from clicking a link to downloading, simulating the resource acquisition process in search results. The first step simulates clicking a link and redirecting the user to the corresponding resource download page. The second step simulates the download interaction on the download page. Throughout the process, page navigation and content changes are monitored. If the page displays a payment option such as "Paid Unlock" or "Recharge Membership," it indicates the presence of a payment step or barrier. If it requires "Enter Account and Password to Log In" or "Authorize Third-Party Account to Log In," and downloading cannot be triggered without logging in, it indicates the presence of a login step or barrier. If it prompts for account registration before continuing to access the page or download, it indicates the presence of a registration step or barrier.

[0140] You can launch a headless browser (running in headless mode), load the full search results page, parse the elements to locate the download trigger element, call the `click()` function to simulate clicking the download trigger element, and listen for the download event. The pop-up window can be obtained through page event listeners.

[0141] You can use page.content() to get the complete HTML source code of the page, or page.inner_text('body') to extract the visible text of the page, to determine whether the page content has changed.

[0142] You can use page.is_visible() to determine if a login form exists on the page. If it does, then there is a login requirement. Alternatively, if a login pop-up appears after the download is triggered, then there is a login requirement.

[0143] If a registration button or form is detected on the page, or if the page forcibly redirects to the registration page or pops up a registration prompt window after a download is triggered in an unregistered state, then it is determined that there is a registration threshold.

[0144] If any of the payment, registration, or login steps are present, the verification result for the resource acquisition threshold is determined to be "failed"; if none of these steps are present, the verification result for the resource acquisition threshold is determined to be "passed".

[0145] By simulating the complete acquisition process, the system verifies the threshold for resource acquisition, ensuring comprehensive and accurate verification. This guarantees that the resources returned to the user can be obtained directly and without obstacles, improving the smoothness of resource downloads and the user's resource search experience.

[0146] Step S306: For each search result returned by the target search engine or for each search result that passes usability verification, obtain the summary information of that search result.

[0147] For search results returned by the target search engine or search results that have passed usability verification, intent verification is required. During intent verification, a summary of the search results needs to be generated first. This summary can be generated using any method; to save costs, it can be generated using the aforementioned large language model.

[0148] The raw information of the search results, including the title, resource description text, and resource metadata, can be input into a large language model, which will then generate a summary of the search results. The large language model can generate the summary information guided by designed summary generation prompts.

[0149] Step S307: Match and verify the summary information with the requested resource type to obtain the verification result of intent verification.

[0150] Intent verification is specifically used to verify the degree of matching between the summary information and the requested resource type, such as similarity. If the matching degree is high or the result of the matching verification is a match, then the intent verification is passed.

[0151] Specifically, the similarity between the summary information and the requested resource type, or between the requested resource type and the requested resource entity, such as cosine similarity, can be used to determine whether the search results pass intent verification.

[0152] For example, if a user's resource search request is "download the open-source code of the paper", the requested resource type is code, and the summary information of the search results only contains the paper's abstract, then it is determined that the two do not match, and the search result fails the intent verification.

[0153] The summary information and the requested resource type can be input into the large language model, so that the large language model can output the verification result of intent verification under the guidance of content judgment prompts.

[0154] Optionally, the summary information and the requested resource type are matched and verified, including: obtaining a structured resource search memory; the structured resource search memory is generated based on the summary information of historical search results that have passed intent verification and their corresponding requested resource types; based on the structured resource search memory, it is determined whether the summary information and the requested resource type have passed the matching verification.

[0155] The structured resource search memory is a structured memory stored in the large language model, which is generated based on historical search results that have passed intent verification.

[0156] During the process of users searching and downloading resources using the large language model, the system records the type of resource requested each time, as well as the summary information of the search results for which the intent verification is successful, adds timestamps, and organizes the data into structured data to obtain structured resource search memory. Each memory in the structured resource search memory represents a set of summary information and requested resource type for intent matching.

[0157] Structured resource search memories can be stored in the form of tables, knowledge graphs, etc.

[0158] After obtaining the user's newly sent resource search request and the summary information of the search results that have passed the resource type and availability verification, the structured resource search memory, the summary information of the search results, and the requested resource type are input into the large language model. The large language model learns the matching rules between the features of the summary information in the structured resource search memory and the requested resource type, thereby realizing the matching verification result between the summary information of the search results obtained in this search and the corresponding requested resource type.

[0159] By using structured memories based on users' historical intent verification, a real and effective personalized benchmark is provided for intent matching verification, avoiding unfounded generalized judgments and improving the accuracy of intent verification. By dynamically associating memories with users' historical download behavior, intent verification is made more aligned with users' long-term needs, reducing misjudgments.

[0160] Step S308: If there are verified search results, generate a response to the resource search request based on the verified search results.

[0161] If all search results returned by the target search engines fail the intent verification, or if all search results returned by the target search engines fail the intent verification within the longest waiting time, it indicates that there are no matching resources for the current requested resource type within the existing search scope. To avoid directly reporting "no search results," a re-search is required. To achieve this re-search, the intent recognition prompts need to be modified. The requested resource type can be directly deleted from the multiple resource types provided to guide the determination of the resource type in the intent recognition prompts. This guides the large language model to determine the resource type of the requested resource entity from the remaining resource types, resulting in a new requested resource type. The target search engine is then re-determined and the retrieval is performed.

[0162] For example, the intent recognition prompt during the initial intent recognition includes a section that reads, "Determine whether the user's intent is related to Python documentation, C++ documentation, JPG images, etc.; if related, return the relevant type." If the requested resource type is Python documentation, then the Python documentation is removed. In the updated intent recognition prompt, this section is updated to read, "Determine whether the user's intent is related to C++ documentation, JPG images, etc.; if related, return the relevant type."

[0163] In step S309, if no search results pass the intent verification, the requested resource type in the multiple resource types in the intent recognition prompt is replaced or deleted to obtain an updated intent recognition prompt. Then, the process returns to the step of performing intent recognition on the user-input resource search request.

[0164] When re-performing intent recognition, i.e., returning to the step of performing intent recognition, the prompt word used is the updated intent recognition prompt word.

[0165] To enable re-searching, the intent-based prompts need to be modified by replacing the requested resource type in the intent-based prompts with other similar types. If a similar type of the requested resource type exists in the intent-based prompts, the requested resource type in the intent-based prompts is directly deleted.

[0166] For example, before the update, the resource types included in the intent recognition prompt were Type 1, Type 2, Type 3, Type 4, and Type 5, where Type 1 is the requested resource type, and Type 5 and Type 6 are similar types to Type 1. The updated intent recognition prompt can include resource types of Type 6, Type 2, Type 3, Type 4, and Type 5, or Type 2, Type 3, Type 4, and Type 5.

[0167] Optionally, the requested resource type in the multiple resource types in the intent recognition prompt can be replaced or deleted to obtain an updated intent recognition prompt, including: determining the similarity type of the requested resource type based on the first knowledge graph representing the similarity of resource types; and replacing the requested resource type in the multiple resource types in the intent recognition prompt with the similarity type of the requested resource type to obtain the updated intent recognition prompt.

[0168] In the first knowledge graph, each resource type is represented as a node, and semantically related nodes are connected by edges. The attributes of the edges are used to characterize the similarity between the two resource types represented by the nodes. The similarity types of the requested resource types can be one or more.

[0169] The accuracy of similarity type determination can be improved by using the first knowledge graph to identify similar resource types. First, locate the node corresponding to the requested resource type in the first knowledge graph, denoted as the target node. Then, determine the similarity type from the resource types corresponding to the nodes connected to the target node.

[0170] You can filter the nodes connected to the target node to find the resource types corresponding to nodes whose edge attributes connected to the target node are greater than a preset threshold, and use these filtered resource types as the similarity types of the requested resource type. Alternatively, you can directly determine the resource type corresponding to the node with the highest attribute of the edge connected to the target node as the similarity type of the requested resource type.

[0171] The nodes connected to the target can be sorted in descending order of the attributes of the connected edges. The resource types corresponding to the top N nodes in the sorted results are taken as the similarity types of the requested resource type. N is a positive integer.

[0172] If no search results pass the availability verification, a search failure response is generated to inform the user that no downloadable resources were found.

[0173] In this embodiment, availability verification of search results ensures that the resources in the search results are available and valid. Matching the summary information of the search results with the requested resource type verifies the intent of the search results, ensuring that the resources provided in the search results are consistent with the resource type corresponding to the user's search intent. These two verifications ensure that the resources in the response are available and accurate, avoiding secondary searches for the user and significantly improving resource download efficiency and search experience. When all search results fail intent verification, the intent recognition prompts are adjusted, and intent recognition and subsequent steps are performed again to achieve re-retrieval and verification of the search results, further improving search coverage and success rate.

[0174] Figure 4 A flowchart illustrating a resource search method provided in another embodiment of this application is shown below. Figure 4 As shown, the resource search method in this embodiment mainly includes the following steps:

[0175] Get user input, such as "Download the TensorFlow documentation for Python";

[0176] Guided by intent recognition prompts, the search intent entered by the user is identified through a large language model, resulting in IP and IP category; where IP represents the aforementioned requested resource entity, and IP category represents the aforementioned requested resource category; for example, "TensorFlow" and "Python documentation";

[0177] Using external tools, select multiple target search engines based on IP classification, and supplement search terms; for example, search engine A and search engine B, and "Python pdf document";

[0178] Information is compiled to obtain IP addresses, IP categories, multiple target search engines, and supplementary search terms;

[0179] Use the IP address and supplementary search terms to search the target search engine;

[0180] Using external tools, the interface of the target search engine is called to perform a search, collect the search results, and perform summary extraction and resource sniffing on the search results; for example, if a PDF download link is found in the third search result of search engine A, the summary information of the third search result and its download link are returned; it is also possible to continue to extract summaries and sniff resources from other search results; among them, resource sniffing is used to verify the availability of resources in the search results;

[0181] Information is aggregated to obtain IP addresses, IP classifications, multiple target search engines, supplementary search terms, summary information of search results, and resource sniffing results;

[0182] Guided by content judgment prompts, the large language model uses the summary information of the search results and IP classification to determine the intent matching of the search results, that is, to determine whether the search results match the search intent.

[0183] If no match is found, modify the intent recognition prompt, re-perform intent recognition, and proceed with subsequent steps. You can replace the IP category in the intent recognition prompt with a similar category, or delete the IP category from the intent recognition prompt.

[0184] If a match is found, a response is generated based on the summarized information and returned to the user. The response returned to the user may include: IP address, IP category, and download link. For example, "TensorFlow", "Python documentation", and URL1.

[0185] Steps not mentioned for using external tools can all be executed by the large language model. The large language model achieves the corresponding functions by calling external tools.

[0186] Corresponding to the resource search method provided in the foregoing embodiments, this application also provides a resource search device. Figure 5 This is a schematic diagram of the structure of the resource search device provided in the embodiments of this application, as shown below. Figure 5As shown, the resource search device includes: an intent recognition module, used to recognize the intent of a user's input resource search request based on a pre-trained large language model, and obtain the requested resource entity and the requested resource type; a search engine determination module, used to determine multiple target search engines from multiple search engines based on the requested resource type, and to expand the search terms based on the requested resource type; a search module, used to call each target search engine to perform a search based on the requested resource entity and the expanded search terms; and a verification and response generation module, used to verify the search results returned by the target search engines, and to generate a response to the resource search request based on the verified search results.

[0187] In one possible implementation, the verification and response generation module includes: a verification unit for performing intent verification and / or availability verification on the search results returned by the target search engine; intent verification is used to verify whether the search results match the requested resource type; and a response generation unit for generating a response to the resource search request based on the verified search results.

[0188] In one possible implementation, the verification unit, during intent verification, specifically performs the following: for each search result returned by the target search engine, obtain summary information of the search results; and perform matching verification between the summary information and the requested resource type to obtain the verification result of intent verification.

[0189] In one possible implementation, the verification unit, during intent verification, is specifically configured to: acquire a structured resource search memory; generate the structured resource search memory based on the summary information of historical search results that have passed intent verification and their corresponding requested resource types; acquire the summary information of each search result returned by the target search engine; and determine, based on the structured resource search memory, whether the summary information and the requested resource type pass the matching verification.

[0190] In one possible implementation, the verification unit, during availability verification, specifically performs the following: verifies the resource acquisition threshold of the search results; if the resource access entry in the search results is a static link, verifies the validity of the link in the search results; if the resource access entry in the search results is hidden in dynamically loaded content, parses the dynamically loaded content, extracts the hidden resource links contained in the dynamically loaded content, and verifies the validity of the hidden resource links.

[0191] In one possible implementation, the verification unit, when performing resource acquisition threshold verification on the search results, is specifically used to: when the link in the search results is valid, simulate the resource acquisition process in the search results, detect whether there is a payment step, a registration step, or a login step, and obtain the verification result of the resource acquisition threshold verification.

[0192] In one possible implementation, the verification unit includes: an availability verification subunit, used to perform availability verification on the search results returned by the target search engine; and an intent verification subunit, used to obtain summary information of each search result returned by the target search engine or each search result that has passed availability verification, and to perform matching verification between the summary information and the requested resource type to obtain the intent verification result.

[0193] In one possible implementation, the intent verification subunit is specifically configured to: obtain summary information of search results that have passed availability verification; obtain a structured resource search memory; generate the structured resource search memory using summary information of historical search results that have passed intent verification and their corresponding requested resource types; and determine, based on the structured resource search memory, whether the summary information and the requested resource type pass the matching verification.

[0194] In one possible implementation, availability verification includes resource access threshold verification and validity verification. The availability verification subunit is specifically used for: performing resource access threshold verification on the search results; and performing validity verification on the links in the search results if the resource access entry in the search results is a static link.

[0195] In one possible implementation, the availability verification subunit is further configured to: if the resource access entry in the search results is hidden in dynamically loaded content, parse the dynamically loaded content, extract the hidden resource links contained in the dynamically loaded content, and verify the validity of the hidden resource links.

[0196] In one possible implementation, the availability verification subunit, when performing resource acquisition threshold verification, specifically performs the following: when the link in the search results is valid, it simulates the resource acquisition process in the search results to detect whether there is a payment step, a registration step, or a login step, and obtains the verification result of the resource acquisition threshold verification.

[0197] In one possible implementation, the intent recognition prompt is used to guide the large language model to extract the requested resource entity of the input resource search request and to determine the requested resource type from multiple resource types; the device also includes a prompt adjustment module, used to: if there is no search result that passes intent verification, replace or delete the requested resource type from the multiple resource types in the intent recognition prompt to obtain an updated intent recognition prompt, and return to the step of performing intent recognition on the user's input resource search request, wherein the prompt used during intent recognition is the updated intent recognition prompt.

[0198] In one possible implementation, the prompt word adjustment module is specifically used to: determine the similarity type of the requested resource type based on the first knowledge graph representing the similarity of resource types, replace the requested resource type among the multiple resource types in the intent recognition prompt words with the similarity type of the requested resource type, and obtain the updated intent recognition prompt words.

[0199] In one possible implementation, the intent recognition module is specifically configured to: based on a pre-trained large language model and guided by intent recognition prompts, perform the following steps: extract entities from the user-input resource search request to obtain the requested resource entity and the resource type entity; if the resource type entity includes polysemous words, obtain the context of the resource search request; extract scene features of the context; and based on the scene features and a second knowledge graph, disambiguate the resource type entity to obtain the requested resource type; wherein the second knowledge graph is used to characterize the association between each resource type entity and various scene features.

[0200] In one possible implementation, the intent recognition module is specifically used to: acquire the user's search intent time-series data; the search intent time-series data is used to record the requested resource types and their associated data corresponding to each search in the user's historical time; input the search intent time-series data and resource search requests into a large language model, so that the large language model performs intent recognition on the resource search requests based on the search intent time-series data, and obtains the requested resource entity and the requested resource type.

[0201] In one possible implementation, the search engine determination module, when determining multiple target search engines, is specifically configured to: obtain multiple candidate search engines corresponding to the requested resource type; obtain the frequency of the user downloading the requested resource type of resource through the search results returned by each candidate search engine; and determine multiple target search engines from the multiple candidate search engines based on the frequency.

[0202] The resource search device provided in this embodiment can execute the resource search method provided in any of the above embodiments. Its implementation principle and technical effect are similar, and will not be described in detail here.

[0203] Figure 6 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Figure 6 As shown, the electronic device 60 provided in this embodiment includes at least one processor 601 and a memory 602. Optionally, the electronic device 60 further includes a communication component 603. The processor 601, memory 602, and communication component 603 are connected via a bus 604.

[0204] In a specific implementation, at least one processor 601 executes computer execution instructions stored in memory 602, causing at least one processor 601 to perform the above-described method.

[0205] The specific implementation process of processor 601 can be found in the above method embodiments, and its implementation principle and technical effect are similar. It will not be repeated here.

[0206] In the above embodiments, it should be understood that the processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in this invention can be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules within the processor.

[0207] The memory may include random access memory (RAM) and may also include non-volatile memory (NVM), such as at least one disk storage device.

[0208] The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. For ease of illustration, the buses shown in the accompanying drawings are not limited to a single bus or a single type of bus.

[0209] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the above-described method.

[0210] This application also provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the above-described method.

[0211] The aforementioned readable storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The readable storage medium can be any available medium accessible to a general-purpose or special-purpose computer.

[0212] An exemplary readable storage medium is coupled to a processor, enabling the processor to read information from and write information to the readable storage medium. Of course, the readable storage medium can also be a component of the processor. The processor and the readable storage medium can reside in an Application Specific Integrated Circuit (ASIC). Alternatively, the processor and the readable storage medium can exist as discrete components in the device.

[0213] The division of units is merely a logical functional division; in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces, devices, or units, and may be electrical, mechanical, or other forms.

[0214] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0215] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0216] If a function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0217] Those skilled in the art will understand that all or part of the steps of the above-described method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above-described method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0218] Finally, it should be noted that other embodiments of the invention will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This invention is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not disclosed herein, and is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope.

Claims

1. A resource search method, characterized in that, include: Based on a pre-trained large language model, the intent of the user's input resource search request is identified to obtain the requested resource entity and the requested resource type. Based on the requested resource type, multiple target search engines are determined from multiple search engines, and the search terms are expanded based on the requested resource type; Based on the requested resource entity and the expanded search terms, each of the target search engines is invoked to perform a search; The search results returned by the target search engine are verified, and a response to the resource search request is generated based on the verified search results.

2. The method according to claim 1, characterized in that, Verification of the search results returned by the target search engine includes: The intent verification is performed on the search results returned by the target search engine; the intent verification is used to verify whether the search results match the requested resource type.

3. The method according to claim 2, characterized in that, Intent verification is performed on the search results returned by the target search engine, including: For each search result returned by the target search engine, obtain summary information of the search result; The summary information is matched and verified with the requested resource type to obtain the verification result of the intent verification.

4. The method according to claim 3, characterized in that, The step of matching and verifying the summary information with the requested resource type includes: Obtain a structured resource search memory; the structured resource search memory is generated based on the summary information of historical search results verified by the intent and their corresponding requested resource types; Based on the structured resource search memory, determine whether the summary information and the requested resource type pass the matching verification.

5. The method according to claim 1, characterized in that, Verification of the search results returned by the target search engine includes: The search results returned by the target search engine are subjected to usability verification; wherein, the usability verification includes resource acquisition threshold verification and validity verification; The usability verification of the search results returned by the target search engine includes: The search results are then subjected to resource acquisition threshold verification. If the resource access entry in the search results is a static link, then the validity of the link in the search results is verified.

6. The method according to claim 5, characterized in that, The usability verification of the search results returned by the target search engine further includes: If the resource access entry in the search results is hidden in dynamically loaded content, then the dynamically loaded content is parsed to extract the hidden resource links contained in the dynamically loaded content. The validity of the hidden resource links is verified.

7. The method according to claim 5 or 6, characterized in that, The verification of resource acquisition thresholds for the search results includes: When the link in the search results is valid, the resource acquisition process in the search results is simulated to detect whether there is a payment step, a registration step, or a login step, and the verification result of the resource acquisition threshold verification is obtained.

8. The method according to any one of claims 1-6, characterized in that, Guided by intent recognition prompts, the large language model performs intent recognition on the input resource search request to obtain the requested resource entity and the requested resource type. The intent recognition prompts include multiple resource types, which are used to guide the large language model to extract the requested resource entity of the input resource search request and to determine the requested resource type from multiple resource types. The method further includes: If no search results pass the intent verification, the requested resource type among the various resource types in the intent recognition prompt is replaced or deleted to obtain an updated intent recognition prompt. The process then returns to the step of performing intent recognition on the user's input resource search request, wherein the prompt used during intent recognition is the updated intent recognition prompt.

9. The method according to claim 8, characterized in that, The process of replacing or deleting the requested resource type from among the various resource types in the intent recognition prompt to obtain an updated intent recognition prompt includes: Based on the first knowledge graph representing the similarity of resource types, the similarity type of the requested resource type is determined; The requested resource type in the multiple resource types in the intent recognition prompt is replaced with a similar type of the requested resource type to obtain the updated intent recognition prompt.

10. The method according to any one of claims 1-6, characterized in that, The pre-trained large language model performs intent recognition on the user's input resource search request to obtain the requested resource entity and the requested resource type, including: Based on a pre-trained large language model, guided by intent recognition prompts, the following steps are performed: Entity extraction is performed on the resource search request input by the user to obtain the requested resource entity and the resource type entity; If the resource type entity includes polysemous words, obtain the context of the resource search request; Extract scene features from the context; Based on the scene features and the second knowledge graph, the resource type entities are disambiguated to obtain the requested resource type; wherein, the second knowledge graph is used to characterize the association between each resource type entity and various scene features.

11. A resource search device, characterized in that, include: The intent recognition module is used to recognize the intent of the user's resource search request based on a pre-trained large language model, and obtain the requested resource entity and the requested resource type. The search engine determination module is used to determine multiple target search engines from multiple search engines based on the requested resource type, and to expand the search terms based on the requested resource type; The search module is used to call each of the target search engines to perform a search based on the requested resource entity and the expanded search terms; The verification and response generation module is used to verify the search results returned by the target search engine and generate a response to the resource search request based on the verified search results.

12. An electronic device, characterized in that, include: Memory, processor; The memory stores computer-executed instructions; The processor executes computer execution instructions stored in the memory, causing the processor to perform the method as described in any one of claims 1-10.

13. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, are used to implement the method as described in any one of claims 1-10.

14. A computer program product, characterized in that, Includes a computer program that, when executed by a processor, implements the method according to any one of claims 1-10.

Images